xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact.c (revision 5b8514eb5c47e8e3689c8fb9dc9fe3f0f65ee1ea)
1 
2 #include "src/mat/impls/baij/seq/baij.h"
3 #include "src/mat/impls/sbaij/seq/sbaij.h"
4 #include "src/inline/ilu.h"
5 #include "include/petscis.h"
6 
7 #if !defined(PETSC_USE_COMPLEX)
8 /*
9   input:
10    F -- numeric factor
11   output:
12    nneg, nzero, npos: matrix inertia
13 */
14 
15 #undef __FUNCT__
16 #define __FUNCT__ "MatGetInertia_SeqSBAIJ"
17 int MatGetInertia_SeqSBAIJ(Mat F,int *nneig,int *nzero,int *npos)
18 {
19   Mat_SeqSBAIJ *fact_ptr = (Mat_SeqSBAIJ*)F->data;
20   PetscScalar  *dd = fact_ptr->a;
21   int          mbs=fact_ptr->mbs,bs=fact_ptr->bs,i,nneig_tmp,npos_tmp,
22                *fi = fact_ptr->i;
23 
24   PetscFunctionBegin;
25   if (bs != 1) SETERRQ1(PETSC_ERR_SUP,"No support for bs: %d >1 yet",bs);
26   nneig_tmp = 0; npos_tmp = 0;
27   for (i=0; i<mbs; i++){
28     if (PetscRealPart(dd[*fi]) > 0.0){
29       npos_tmp++;
30     } else if (PetscRealPart(dd[*fi]) < 0.0){
31       nneig_tmp++;
32     }
33     fi++;
34   }
35   if (nneig) *nneig = nneig_tmp;
36   if (npos)  *npos  = npos_tmp;
37   if (nzero) *nzero = mbs - nneig_tmp - npos_tmp;
38 
39   PetscFunctionReturn(0);
40 }
41 #endif /* !defined(PETSC_USE_COMPLEX) */
42 
43 /* Using Modified Sparse Row (MSR) storage.
44 See page 85, "Iterative Methods ..." by Saad. */
45 /*
46     Symbolic U^T*D*U factorization for SBAIJ format. Modified from SSF of YSMP.
47 */
48 /* Use Modified Sparse Row storage for u and ju, see Saad pp.85 */
49 #undef __FUNCT__
50 #define __FUNCT__ "MatCholeskyFactorSymbolic_SeqSBAIJ"
51 int MatCholeskyFactorSymbolic_SeqSBAIJ(Mat A,IS perm,MatFactorInfo *info,Mat *B)
52 {
53   Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b;
54   int          *rip,ierr,i,mbs = a->mbs,*ai,*aj;
55   int          *jutmp,bs = a->bs,bs2=a->bs2;
56   int          m,realloc = 0,prow;
57   int          *jl,*q,jmin,jmax,juidx,nzk,qm,*iu,*ju,k,j,vj,umax,maxadd;
58   int          *il,ili,nextprow;
59   PetscReal    f = info->fill;
60   PetscTruth   perm_identity;
61 
62   PetscFunctionBegin;
63   /* check whether perm is the identity mapping */
64   ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr);
65 
66   /* -- inplace factorization, i.e., use sbaij for *B -- */
67   if (perm_identity && bs==1 ){
68     if (!perm_identity) a->permute = PETSC_TRUE;
69 
70   ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr);
71 
72   if (perm_identity){ /* without permutation */
73     ai = a->i; aj = a->j;
74   } else {            /* non-trivial permutation */
75     ierr = MatReorderingSeqSBAIJ(A,perm);CHKERRQ(ierr);
76     ai = a->inew; aj = a->jnew;
77   }
78 
79   /* initialization */
80   ierr  = PetscMalloc((mbs+1)*sizeof(int),&iu);CHKERRQ(ierr);
81   umax  = (int)(f*ai[mbs] + 1);
82   ierr  = PetscMalloc(umax*sizeof(int),&ju);CHKERRQ(ierr);
83   iu[0] = 0;
84   juidx = 0; /* index for ju */
85   ierr  = PetscMalloc((3*mbs+1)*sizeof(int),&jl);CHKERRQ(ierr); /* linked list for getting pivot row */
86   q     = jl + mbs;   /* linked list for col index of active row */
87   il    = q  + mbs;
88   for (i=0; i<mbs; i++){
89     jl[i] = mbs;
90     q[i]  = 0;
91     il[i] = 0;
92   }
93 
94   /* for each row k */
95   for (k=0; k<mbs; k++){
96     nzk  = 0; /* num. of nz blocks in k-th block row with diagonal block excluded */
97     q[k] = mbs;
98     /* initialize nonzero structure of k-th row to row rip[k] of A */
99     jmin = ai[rip[k]] +1; /* exclude diag[k] */
100     jmax = ai[rip[k]+1];
101     for (j=jmin; j<jmax; j++){
102       vj = rip[aj[j]]; /* col. value */
103       if(vj > k){
104         qm = k;
105         do {
106           m  = qm; qm = q[m];
107         } while(qm < vj);
108         if (qm == vj) {
109           SETERRQ(1," error: duplicate entry in A\n");
110         }
111         nzk++;
112         q[m]  = vj;
113         q[vj] = qm;
114       } /* if(vj > k) */
115     } /* for (j=jmin; j<jmax; j++) */
116 
117     /* modify nonzero structure of k-th row by computing fill-in
118        for each row i to be merged in */
119     prow = k;
120     prow = jl[prow]; /* next pivot row (== mbs for symbolic factorization) */
121 
122     while (prow < k){
123       nextprow = jl[prow];
124 
125       /* merge row prow into k-th row */
126       ili = il[prow];
127       jmin = ili + 1;  /* points to 2nd nzero entry in U(prow,k:mbs-1) */
128       jmax = iu[prow+1];
129       qm = k;
130       for (j=jmin; j<jmax; j++){
131         vj = ju[j];
132         do {
133           m = qm; qm = q[m];
134         } while (qm < vj);
135         if (qm != vj){  /* a fill */
136           nzk++; q[m] = vj; q[vj] = qm; qm = vj;
137         }
138       } /* end of for (j=jmin; j<jmax; j++) */
139       if (jmin < jmax){
140         il[prow] = jmin;
141         j = ju[jmin];
142         jl[prow] = jl[j]; jl[j] = prow;  /* update jl */
143       }
144       prow = nextprow;
145     }
146 
147     /* update il and jl */
148     if (nzk > 0){
149       i = q[k]; /* col value of the first nonzero element in U(k, k+1:mbs-1) */
150       jl[k] = jl[i]; jl[i] = k;
151       il[k] = iu[k] + 1;
152     }
153     iu[k+1] = iu[k] + nzk + 1;  /* include diag[k] */
154 
155     /* allocate more space to ju if needed */
156     if (iu[k+1] > umax) {
157       /* estimate how much additional space we will need */
158       /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */
159       /* just double the memory each time */
160       maxadd = umax;
161       if (maxadd < nzk) maxadd = (mbs-k)*(nzk+1)/2;
162       umax += maxadd;
163 
164       /* allocate a longer ju */
165       ierr = PetscMalloc(umax*sizeof(int),&jutmp);CHKERRQ(ierr);
166       ierr = PetscMemcpy(jutmp,ju,iu[k]*sizeof(int));CHKERRQ(ierr);
167       ierr = PetscFree(ju);CHKERRQ(ierr);
168       ju   = jutmp;
169       realloc++; /* count how many times we realloc */
170     }
171 
172     /* save nonzero structure of k-th row in ju */
173     ju[juidx++] = k; /* diag[k] */
174     i = k;
175     while (nzk --) {
176       i           = q[i];
177       ju[juidx++] = i;
178     }
179   }
180 
181   if (ai[mbs] != 0) {
182     PetscReal af = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]);
183     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Reallocs %d Fill ratio:given %g needed %g\n",realloc,f,af);
184     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Run with -pc_cholesky_fill %g or use \n",af);
185     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:PCCholeskySetFill(pc,%g);\n",af);
186     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:for best performance.\n");
187   } else {
188      PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Empty matrix.\n");
189   }
190 
191   ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr);
192   /* ierr = PetscFree(q);CHKERRQ(ierr); */
193   ierr = PetscFree(jl);CHKERRQ(ierr);
194 
195   /* put together the new matrix */
196   ierr = MatCreate(A->comm,bs*mbs,bs*mbs,bs*mbs,bs*mbs,B);CHKERRQ(ierr);
197   ierr = MatSetType(*B,A->type_name);CHKERRQ(ierr);
198   ierr = MatSeqSBAIJSetPreallocation(*B,bs,0,PETSC_NULL);CHKERRQ(ierr);
199 
200   /* PetscLogObjectParent(*B,iperm); */
201   b = (Mat_SeqSBAIJ*)(*B)->data;
202   ierr = PetscFree(b->imax);CHKERRQ(ierr);
203   b->singlemalloc = PETSC_FALSE;
204   /* the next line frees the default space generated by the Create() */
205   ierr = PetscFree(b->a);CHKERRQ(ierr);
206   ierr = PetscFree(b->ilen);CHKERRQ(ierr);
207   ierr = PetscMalloc((iu[mbs]+1)*sizeof(MatScalar)*bs2,&b->a);CHKERRQ(ierr);
208   b->j    = ju;
209   b->i    = iu;
210   b->diag = 0;
211   b->ilen = 0;
212   b->imax = 0;
213   b->row  = perm;
214   b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */
215   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
216   b->icol = perm;
217   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
218   ierr    = PetscMalloc((bs*mbs+bs)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
219   /* In b structure:  Free imax, ilen, old a, old j.
220      Allocate idnew, solve_work, new a, new j */
221   PetscLogObjectMemory(*B,(iu[mbs]-mbs)*(sizeof(int)+sizeof(MatScalar)));
222   b->maxnz = b->nz = iu[mbs];
223 
224   (*B)->factor                 = FACTOR_CHOLESKY;
225   (*B)->info.factor_mallocs    = realloc;
226   (*B)->info.fill_ratio_given  = f;
227   if (ai[mbs] != 0) {
228     (*B)->info.fill_ratio_needed = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]);
229   } else {
230     (*B)->info.fill_ratio_needed = 0.0;
231   }
232 
233 
234   (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering;
235   (*B)->ops->solve           = MatSolve_SeqSBAIJ_1_NaturalOrdering;
236   PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=1\n");
237 
238   PetscFunctionReturn(0);
239   }
240   /* -----------  end of new code --------------------*/
241 
242 
243   if (!perm_identity) a->permute = PETSC_TRUE;
244 
245   ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr);
246 
247   if (perm_identity){ /* without permutation */
248     ai = a->i; aj = a->j;
249   } else {            /* non-trivial permutation */
250     ierr = MatReorderingSeqSBAIJ(A,perm);CHKERRQ(ierr);
251     ai = a->inew; aj = a->jnew;
252   }
253 
254   /* initialization */
255   ierr  = PetscMalloc((mbs+1)*sizeof(int),&iu);CHKERRQ(ierr);
256   umax  = (int)(f*ai[mbs] + 1); umax += mbs + 1;
257   ierr  = PetscMalloc(umax*sizeof(int),&ju);CHKERRQ(ierr);
258   iu[0] = mbs+1;
259   juidx = mbs + 1; /* index for ju */
260   ierr  = PetscMalloc(2*mbs*sizeof(int),&jl);CHKERRQ(ierr); /* linked list for pivot row */
261   q     = jl + mbs;   /* linked list for col index */
262   for (i=0; i<mbs; i++){
263     jl[i] = mbs;
264     q[i] = 0;
265   }
266 
267   /* for each row k */
268   for (k=0; k<mbs; k++){
269     for (i=0; i<mbs; i++) q[i] = 0;  /* to be removed! */
270     nzk  = 0; /* num. of nz blocks in k-th block row with diagonal block excluded */
271     q[k] = mbs;
272     /* initialize nonzero structure of k-th row to row rip[k] of A */
273     jmin = ai[rip[k]] +1; /* exclude diag[k] */
274     jmax = ai[rip[k]+1];
275     for (j=jmin; j<jmax; j++){
276       vj = rip[aj[j]]; /* col. value */
277       if(vj > k){
278         qm = k;
279         do {
280           m  = qm; qm = q[m];
281         } while(qm < vj);
282         if (qm == vj) {
283           SETERRQ(1," error: duplicate entry in A\n");
284         }
285         nzk++;
286         q[m]  = vj;
287         q[vj] = qm;
288       } /* if(vj > k) */
289     } /* for (j=jmin; j<jmax; j++) */
290 
291     /* modify nonzero structure of k-th row by computing fill-in
292        for each row i to be merged in */
293     prow = k;
294     prow = jl[prow]; /* next pivot row (== mbs for symbolic factorization) */
295 
296     while (prow < k){
297       /* merge row prow into k-th row */
298       jmin = iu[prow] + 1; jmax = iu[prow+1];
299       qm = k;
300       for (j=jmin; j<jmax; j++){
301         vj = ju[j];
302         do {
303           m = qm; qm = q[m];
304         } while (qm < vj);
305         if (qm != vj){
306          nzk++; q[m] = vj; q[vj] = qm; qm = vj;
307         }
308       }
309       prow = jl[prow]; /* next pivot row */
310     }
311 
312     /* add k to row list for first nonzero element in k-th row */
313     if (nzk > 0){
314       i = q[k]; /* col value of first nonzero element in U(k, k+1:mbs-1) */
315       jl[k] = jl[i]; jl[i] = k;
316     }
317     iu[k+1] = iu[k] + nzk;
318 
319     /* allocate more space to ju if needed */
320     if (iu[k+1] > umax) {
321       /* estimate how much additional space we will need */
322       /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */
323       /* just double the memory each time */
324       maxadd = umax;
325       if (maxadd < nzk) maxadd = (mbs-k)*(nzk+1)/2;
326       umax += maxadd;
327 
328       /* allocate a longer ju */
329       ierr = PetscMalloc(umax*sizeof(int),&jutmp);CHKERRQ(ierr);
330       ierr = PetscMemcpy(jutmp,ju,iu[k]*sizeof(int));CHKERRQ(ierr);
331       ierr = PetscFree(ju);CHKERRQ(ierr);
332       ju   = jutmp;
333       realloc++; /* count how many times we realloc */
334     }
335 
336     /* save nonzero structure of k-th row in ju */
337     i=k;
338     while (nzk --) {
339       i           = q[i];
340       ju[juidx++] = i;
341     }
342   }
343 
344   if (ai[mbs] != 0) {
345     PetscReal af = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]);
346     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Reallocs %d Fill ratio:given %g needed %g\n",realloc,f,af);
347     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Run with -pc_cholesky_fill %g or use \n",af);
348     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:PCCholeskySetFill(pc,%g);\n",af);
349     PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:for best performance.\n");
350   } else {
351      PetscLogInfo(A,"MatCholeskyFactorSymbolic_SeqSBAIJ:Empty matrix.\n");
352   }
353 
354   ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr);
355   /* ierr = PetscFree(q);CHKERRQ(ierr); */
356   ierr = PetscFree(jl);CHKERRQ(ierr);
357 
358   /* put together the new matrix */
359   ierr = MatCreate(A->comm,bs*mbs,bs*mbs,bs*mbs,bs*mbs,B);CHKERRQ(ierr);
360   ierr = MatSetType(*B,A->type_name);CHKERRQ(ierr);
361   ierr = MatSeqSBAIJSetPreallocation(*B,bs,0,PETSC_NULL);CHKERRQ(ierr);
362 
363   /* PetscLogObjectParent(*B,iperm); */
364   b = (Mat_SeqSBAIJ*)(*B)->data;
365   ierr = PetscFree(b->imax);CHKERRQ(ierr);
366   b->singlemalloc = PETSC_FALSE;
367   /* the next line frees the default space generated by the Create() */
368   ierr = PetscFree(b->a);CHKERRQ(ierr);
369   ierr = PetscFree(b->ilen);CHKERRQ(ierr);
370   ierr = PetscMalloc((iu[mbs]+1)*sizeof(MatScalar)*bs2,&b->a);CHKERRQ(ierr);
371   b->j    = ju;
372   b->i    = iu;
373   b->diag = 0;
374   b->ilen = 0;
375   b->imax = 0;
376   b->row  = perm;
377   b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */
378   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
379   b->icol = perm;
380   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
381   ierr    = PetscMalloc((bs*mbs+bs)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
382   /* In b structure:  Free imax, ilen, old a, old j.
383      Allocate idnew, solve_work, new a, new j */
384   PetscLogObjectMemory(*B,(iu[mbs]-mbs)*(sizeof(int)+sizeof(MatScalar)));
385   b->maxnz = b->nz = iu[mbs];
386 
387   (*B)->factor                 = FACTOR_CHOLESKY;
388   (*B)->info.factor_mallocs    = realloc;
389   (*B)->info.fill_ratio_given  = f;
390   if (ai[mbs] != 0) {
391     (*B)->info.fill_ratio_needed = ((PetscReal)iu[mbs])/((PetscReal)ai[mbs]);
392   } else {
393     (*B)->info.fill_ratio_needed = 0.0;
394   }
395 
396   if (perm_identity){
397     switch (bs) {
398       case 1:
399         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering;
400         (*B)->ops->solve           = MatSolve_SeqSBAIJ_1_NaturalOrdering;
401         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=1\n");
402         break;
403       case 2:
404         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering;
405         (*B)->ops->solve           = MatSolve_SeqSBAIJ_2_NaturalOrdering;
406         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=2\n");
407         break;
408       case 3:
409         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering;
410         (*B)->ops->solve           = MatSolve_SeqSBAIJ_3_NaturalOrdering;
411         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:sing special in-place natural ordering factor and solve BS=3\n");
412         break;
413       case 4:
414         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering;
415         (*B)->ops->solve           = MatSolve_SeqSBAIJ_4_NaturalOrdering;
416         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=4\n");
417         break;
418       case 5:
419         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering;
420         (*B)->ops->solve           = MatSolve_SeqSBAIJ_5_NaturalOrdering;
421         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=5\n");
422         break;
423       case 6:
424         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering;
425         (*B)->ops->solve           = MatSolve_SeqSBAIJ_6_NaturalOrdering;
426         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=6\n");
427         break;
428       case 7:
429         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering;
430         (*B)->ops->solve           = MatSolve_SeqSBAIJ_7_NaturalOrdering;
431         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS=7\n");
432       break;
433       default:
434         (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering;
435         (*B)->ops->solve           = MatSolve_SeqSBAIJ_N_NaturalOrdering;
436         PetscLogInfo(A,"MatICCFactorSymbolic_SeqSBAIJ:Using special in-place natural ordering factor and solve BS>7\n");
437       break;
438     }
439   }
440 
441   PetscFunctionReturn(0);
442 }
443 
444 
445 #undef __FUNCT__
446 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_N"
447 int MatCholeskyFactorNumeric_SeqSBAIJ_N(Mat A,Mat *B)
448 {
449   Mat                C = *B;
450   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
451   IS                 perm = b->row;
452   int                *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
453   int                *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
454   int                bs=a->bs,bs2 = a->bs2;
455   MatScalar          *ba = b->a,*aa,*ap,*dk,*uik;
456   MatScalar          *u,*diag,*rtmp,*rtmp_ptr;
457   MatScalar          *work;
458   int                *pivots;
459 
460   PetscFunctionBegin;
461 
462   /* initialization */
463   ierr = PetscMalloc(bs2*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
464   ierr = PetscMemzero(rtmp,bs2*mbs*sizeof(MatScalar));CHKERRQ(ierr);
465   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
466   jl   = il + mbs;
467   for (i=0; i<mbs; i++) {
468     jl[i] = mbs; il[0] = 0;
469   }
470   ierr = PetscMalloc((2*bs2+bs)*sizeof(MatScalar),&dk);CHKERRQ(ierr);
471   uik  = dk + bs2;
472   work = uik + bs2;
473   ierr = PetscMalloc(bs*sizeof(int),&pivots);CHKERRQ(ierr);
474 
475   ierr  = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr);
476 
477   /* check permutation */
478   if (!a->permute){
479     ai = a->i; aj = a->j; aa = a->a;
480   } else {
481     ai   = a->inew; aj = a->jnew;
482     ierr = PetscMalloc(bs2*ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr);
483     ierr = PetscMemcpy(aa,a->a,bs2*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr);
484     ierr = PetscMalloc(ai[mbs]*sizeof(int),&a2anew);CHKERRQ(ierr);
485     ierr = PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr);
486 
487     for (i=0; i<mbs; i++){
488       jmin = ai[i]; jmax = ai[i+1];
489       for (j=jmin; j<jmax; j++){
490         while (a2anew[j] != j){
491           k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k;
492           for (k1=0; k1<bs2; k1++){
493             dk[k1]       = aa[k*bs2+k1];
494             aa[k*bs2+k1] = aa[j*bs2+k1];
495             aa[j*bs2+k1] = dk[k1];
496           }
497         }
498         /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */
499         if (i > aj[j]){
500           /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */
501           ap = aa + j*bs2;                     /* ptr to the beginning of j-th block of aa */
502           for (k=0; k<bs2; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */
503           for (k=0; k<bs; k++){               /* j-th block of aa <- dk^T */
504             for (k1=0; k1<bs; k1++) *ap++ = dk[k + bs*k1];
505           }
506         }
507       }
508     }
509     ierr = PetscFree(a2anew);CHKERRQ(ierr);
510   }
511 
512   /* for each row k */
513   for (k = 0; k<mbs; k++){
514 
515     /*initialize k-th row with elements nonzero in row perm(k) of A */
516     jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1];
517 
518     ap = aa + jmin*bs2;
519     for (j = jmin; j < jmax; j++){
520       vj = perm_ptr[aj[j]];         /* block col. index */
521       rtmp_ptr = rtmp + vj*bs2;
522       for (i=0; i<bs2; i++) *rtmp_ptr++ = *ap++;
523     }
524 
525     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
526     ierr = PetscMemcpy(dk,rtmp+k*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr);
527     i = jl[k]; /* first row to be added to k_th row  */
528 
529     while (i < k){
530       nexti = jl[i]; /* next row to be added to k_th row */
531 
532       /* compute multiplier */
533       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
534 
535       /* uik = -inv(Di)*U_bar(i,k) */
536       diag = ba + i*bs2;
537       u    = ba + ili*bs2;
538       ierr = PetscMemzero(uik,bs2*sizeof(MatScalar));CHKERRQ(ierr);
539       Kernel_A_gets_A_minus_B_times_C(bs,uik,diag,u);
540 
541       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
542       Kernel_A_gets_A_plus_Btranspose_times_C(bs,dk,uik,u);
543 
544       /* update -U(i,k) */
545       ierr = PetscMemcpy(ba+ili*bs2,uik,bs2*sizeof(MatScalar));CHKERRQ(ierr);
546 
547       /* add multiple of row i to k-th row ... */
548       jmin = ili + 1; jmax = bi[i+1];
549       if (jmin < jmax){
550         for (j=jmin; j<jmax; j++) {
551           /* rtmp += -U(i,k)^T * U_bar(i,j) */
552           rtmp_ptr = rtmp + bj[j]*bs2;
553           u = ba + j*bs2;
554           Kernel_A_gets_A_plus_Btranspose_times_C(bs,rtmp_ptr,uik,u);
555         }
556 
557         /* ... add i to row list for next nonzero entry */
558         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
559         j     = bj[jmin];
560         jl[i] = jl[j]; jl[j] = i; /* update jl */
561       }
562       i = nexti;
563     }
564 
565     /* save nonzero entries in k-th row of U ... */
566 
567     /* invert diagonal block */
568     diag = ba+k*bs2;
569     ierr = PetscMemcpy(diag,dk,bs2*sizeof(MatScalar));CHKERRQ(ierr);
570     Kernel_A_gets_inverse_A(bs,diag,pivots,work);
571 
572     jmin = bi[k]; jmax = bi[k+1];
573     if (jmin < jmax) {
574       for (j=jmin; j<jmax; j++){
575          vj = bj[j];           /* block col. index of U */
576          u   = ba + j*bs2;
577          rtmp_ptr = rtmp + vj*bs2;
578          for (k1=0; k1<bs2; k1++){
579            *u++        = *rtmp_ptr;
580            *rtmp_ptr++ = 0.0;
581          }
582       }
583 
584       /* ... add k to row list for first nonzero entry in k-th row */
585       il[k] = jmin;
586       i     = bj[jmin];
587       jl[k] = jl[i]; jl[i] = k;
588     }
589   }
590 
591   ierr = PetscFree(rtmp);CHKERRQ(ierr);
592   ierr = PetscFree(il);CHKERRQ(ierr);
593   ierr = PetscFree(dk);CHKERRQ(ierr);
594   ierr = PetscFree(pivots);CHKERRQ(ierr);
595   if (a->permute){
596     ierr = PetscFree(aa);CHKERRQ(ierr);
597   }
598 
599   ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr);
600   C->factor       = FACTOR_CHOLESKY;
601   C->assembled    = PETSC_TRUE;
602   C->preallocated = PETSC_TRUE;
603   PetscLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */
604   PetscFunctionReturn(0);
605 }
606 
607 #undef __FUNCT__
608 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering"
609 int MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering(Mat A,Mat *B)
610 {
611   Mat                C = *B;
612   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
613   int                ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
614   int                *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
615   int                bs=a->bs,bs2 = a->bs2;
616   MatScalar          *ba = b->a,*aa,*ap,*dk,*uik;
617   MatScalar          *u,*diag,*rtmp,*rtmp_ptr;
618   MatScalar          *work;
619   int                *pivots;
620 
621   PetscFunctionBegin;
622 
623   /* initialization */
624 
625   ierr = PetscMalloc(bs2*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
626   ierr = PetscMemzero(rtmp,bs2*mbs*sizeof(MatScalar));CHKERRQ(ierr);
627   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
628   jl   = il + mbs;
629   for (i=0; i<mbs; i++) {
630     jl[i] = mbs; il[0] = 0;
631   }
632   ierr = PetscMalloc((2*bs2+bs)*sizeof(MatScalar),&dk);CHKERRQ(ierr);
633   uik  = dk + bs2;
634   work = uik + bs2;
635   ierr = PetscMalloc(bs*sizeof(int),&pivots);CHKERRQ(ierr);
636 
637   ai = a->i; aj = a->j; aa = a->a;
638 
639   /* for each row k */
640   for (k = 0; k<mbs; k++){
641 
642     /*initialize k-th row with elements nonzero in row k of A */
643     jmin = ai[k]; jmax = ai[k+1];
644     ap = aa + jmin*bs2;
645     for (j = jmin; j < jmax; j++){
646       vj = aj[j];         /* block col. index */
647       rtmp_ptr = rtmp + vj*bs2;
648       for (i=0; i<bs2; i++) *rtmp_ptr++ = *ap++;
649     }
650 
651     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
652     ierr = PetscMemcpy(dk,rtmp+k*bs2,bs2*sizeof(MatScalar));CHKERRQ(ierr);
653     i = jl[k]; /* first row to be added to k_th row  */
654 
655     while (i < k){
656       nexti = jl[i]; /* next row to be added to k_th row */
657 
658       /* compute multiplier */
659       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
660 
661       /* uik = -inv(Di)*U_bar(i,k) */
662       diag = ba + i*bs2;
663       u    = ba + ili*bs2;
664       ierr = PetscMemzero(uik,bs2*sizeof(MatScalar));CHKERRQ(ierr);
665       Kernel_A_gets_A_minus_B_times_C(bs,uik,diag,u);
666 
667       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
668       Kernel_A_gets_A_plus_Btranspose_times_C(bs,dk,uik,u);
669 
670       /* update -U(i,k) */
671       ierr = PetscMemcpy(ba+ili*bs2,uik,bs2*sizeof(MatScalar));CHKERRQ(ierr);
672 
673       /* add multiple of row i to k-th row ... */
674       jmin = ili + 1; jmax = bi[i+1];
675       if (jmin < jmax){
676         for (j=jmin; j<jmax; j++) {
677           /* rtmp += -U(i,k)^T * U_bar(i,j) */
678           rtmp_ptr = rtmp + bj[j]*bs2;
679           u = ba + j*bs2;
680           Kernel_A_gets_A_plus_Btranspose_times_C(bs,rtmp_ptr,uik,u);
681         }
682 
683         /* ... add i to row list for next nonzero entry */
684         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
685         j     = bj[jmin];
686         jl[i] = jl[j]; jl[j] = i; /* update jl */
687       }
688       i = nexti;
689     }
690 
691     /* save nonzero entries in k-th row of U ... */
692 
693     /* invert diagonal block */
694     diag = ba+k*bs2;
695     ierr = PetscMemcpy(diag,dk,bs2*sizeof(MatScalar));CHKERRQ(ierr);
696     Kernel_A_gets_inverse_A(bs,diag,pivots,work);
697 
698     jmin = bi[k]; jmax = bi[k+1];
699     if (jmin < jmax) {
700       for (j=jmin; j<jmax; j++){
701          vj = bj[j];           /* block col. index of U */
702          u   = ba + j*bs2;
703          rtmp_ptr = rtmp + vj*bs2;
704          for (k1=0; k1<bs2; k1++){
705            *u++        = *rtmp_ptr;
706            *rtmp_ptr++ = 0.0;
707          }
708       }
709 
710       /* ... add k to row list for first nonzero entry in k-th row */
711       il[k] = jmin;
712       i     = bj[jmin];
713       jl[k] = jl[i]; jl[i] = k;
714     }
715   }
716 
717   ierr = PetscFree(rtmp);CHKERRQ(ierr);
718   ierr = PetscFree(il);CHKERRQ(ierr);
719   ierr = PetscFree(dk);CHKERRQ(ierr);
720   ierr = PetscFree(pivots);CHKERRQ(ierr);
721 
722   C->factor    = FACTOR_CHOLESKY;
723   C->assembled = PETSC_TRUE;
724   C->preallocated = PETSC_TRUE;
725   PetscLogFlops(1.3333*bs*bs2*b->mbs); /* from inverting diagonal blocks */
726   PetscFunctionReturn(0);
727 }
728 
729 /*
730     Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP.
731     Version for blocks 2 by 2.
732 */
733 #undef __FUNCT__
734 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_2"
735 int MatCholeskyFactorNumeric_SeqSBAIJ_2(Mat A,Mat *B)
736 {
737   Mat                C = *B;
738   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
739   IS                 perm = b->row;
740   int                *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
741   int                *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
742   MatScalar          *ba = b->a,*aa,*ap,*dk,*uik;
743   MatScalar          *u,*diag,*rtmp,*rtmp_ptr;
744 
745   PetscFunctionBegin;
746 
747   /* initialization */
748   /* il and jl record the first nonzero element in each row of the accessing
749      window U(0:k, k:mbs-1).
750      jl:    list of rows to be added to uneliminated rows
751             i>= k: jl(i) is the first row to be added to row i
752             i<  k: jl(i) is the row following row i in some list of rows
753             jl(i) = mbs indicates the end of a list
754      il(i): points to the first nonzero element in columns k,...,mbs-1 of
755             row i of U */
756   ierr = PetscMalloc(4*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
757   ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr);
758   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
759   jl   = il + mbs;
760   for (i=0; i<mbs; i++) {
761     jl[i] = mbs; il[0] = 0;
762   }
763   ierr = PetscMalloc(8*sizeof(MatScalar),&dk);CHKERRQ(ierr);
764   uik  = dk + 4;
765   ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr);
766 
767   /* check permutation */
768   if (!a->permute){
769     ai = a->i; aj = a->j; aa = a->a;
770   } else {
771     ai   = a->inew; aj = a->jnew;
772     ierr = PetscMalloc(4*ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr);
773     ierr = PetscMemcpy(aa,a->a,4*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr);
774     ierr = PetscMalloc(ai[mbs]*sizeof(int),&a2anew);CHKERRQ(ierr);
775     ierr = PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr);
776 
777     for (i=0; i<mbs; i++){
778       jmin = ai[i]; jmax = ai[i+1];
779       for (j=jmin; j<jmax; j++){
780         while (a2anew[j] != j){
781           k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k;
782           for (k1=0; k1<4; k1++){
783             dk[k1]       = aa[k*4+k1];
784             aa[k*4+k1] = aa[j*4+k1];
785             aa[j*4+k1] = dk[k1];
786           }
787         }
788         /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */
789         if (i > aj[j]){
790           /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */
791           ap = aa + j*4;     /* ptr to the beginning of the block */
792           dk[1] = ap[1];     /* swap ap[1] and ap[2] */
793           ap[1] = ap[2];
794           ap[2] = dk[1];
795         }
796       }
797     }
798     ierr = PetscFree(a2anew);CHKERRQ(ierr);
799   }
800 
801   /* for each row k */
802   for (k = 0; k<mbs; k++){
803 
804     /*initialize k-th row with elements nonzero in row perm(k) of A */
805     jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1];
806     ap = aa + jmin*4;
807     for (j = jmin; j < jmax; j++){
808       vj = perm_ptr[aj[j]];         /* block col. index */
809       rtmp_ptr = rtmp + vj*4;
810       for (i=0; i<4; i++) *rtmp_ptr++ = *ap++;
811     }
812 
813     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
814     ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr);
815     i = jl[k]; /* first row to be added to k_th row  */
816 
817     while (i < k){
818       nexti = jl[i]; /* next row to be added to k_th row */
819 
820       /* compute multiplier */
821       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
822 
823       /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */
824       diag = ba + i*4;
825       u    = ba + ili*4;
826       uik[0] = -(diag[0]*u[0] + diag[2]*u[1]);
827       uik[1] = -(diag[1]*u[0] + diag[3]*u[1]);
828       uik[2] = -(diag[0]*u[2] + diag[2]*u[3]);
829       uik[3] = -(diag[1]*u[2] + diag[3]*u[3]);
830 
831       /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */
832       dk[0] += uik[0]*u[0] + uik[1]*u[1];
833       dk[1] += uik[2]*u[0] + uik[3]*u[1];
834       dk[2] += uik[0]*u[2] + uik[1]*u[3];
835       dk[3] += uik[2]*u[2] + uik[3]*u[3];
836 
837       /* update -U(i,k): ba[ili] = uik */
838       ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr);
839 
840       /* add multiple of row i to k-th row ... */
841       jmin = ili + 1; jmax = bi[i+1];
842       if (jmin < jmax){
843         for (j=jmin; j<jmax; j++) {
844           /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */
845           rtmp_ptr = rtmp + bj[j]*4;
846           u = ba + j*4;
847           rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1];
848           rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1];
849           rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3];
850           rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3];
851         }
852 
853         /* ... add i to row list for next nonzero entry */
854         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
855         j     = bj[jmin];
856         jl[i] = jl[j]; jl[j] = i; /* update jl */
857       }
858       i = nexti;
859     }
860 
861     /* save nonzero entries in k-th row of U ... */
862 
863     /* invert diagonal block */
864     diag = ba+k*4;
865     ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr);
866     ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr);
867 
868     jmin = bi[k]; jmax = bi[k+1];
869     if (jmin < jmax) {
870       for (j=jmin; j<jmax; j++){
871          vj = bj[j];           /* block col. index of U */
872          u   = ba + j*4;
873          rtmp_ptr = rtmp + vj*4;
874          for (k1=0; k1<4; k1++){
875            *u++        = *rtmp_ptr;
876            *rtmp_ptr++ = 0.0;
877          }
878       }
879 
880       /* ... add k to row list for first nonzero entry in k-th row */
881       il[k] = jmin;
882       i     = bj[jmin];
883       jl[k] = jl[i]; jl[i] = k;
884     }
885   }
886 
887   ierr = PetscFree(rtmp);CHKERRQ(ierr);
888   ierr = PetscFree(il);CHKERRQ(ierr);
889   ierr = PetscFree(dk);CHKERRQ(ierr);
890   if (a->permute) {
891     ierr = PetscFree(aa);CHKERRQ(ierr);
892   }
893   ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr);
894   C->factor    = FACTOR_CHOLESKY;
895   C->assembled = PETSC_TRUE;
896   C->preallocated = PETSC_TRUE;
897   PetscLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */
898   PetscFunctionReturn(0);
899 }
900 
901 /*
902       Version for when blocks are 2 by 2 Using natural ordering
903 */
904 #undef __FUNCT__
905 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering"
906 int MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering(Mat A,Mat *B)
907 {
908   Mat                C = *B;
909   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
910   int                ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
911   int                *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
912   MatScalar          *ba = b->a,*aa,*ap,*dk,*uik;
913   MatScalar          *u,*diag,*rtmp,*rtmp_ptr;
914 
915   PetscFunctionBegin;
916 
917   /* initialization */
918   /* il and jl record the first nonzero element in each row of the accessing
919      window U(0:k, k:mbs-1).
920      jl:    list of rows to be added to uneliminated rows
921             i>= k: jl(i) is the first row to be added to row i
922             i<  k: jl(i) is the row following row i in some list of rows
923             jl(i) = mbs indicates the end of a list
924      il(i): points to the first nonzero element in columns k,...,mbs-1 of
925             row i of U */
926   ierr = PetscMalloc(4*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
927   ierr = PetscMemzero(rtmp,4*mbs*sizeof(MatScalar));CHKERRQ(ierr);
928   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
929   jl   = il + mbs;
930   for (i=0; i<mbs; i++) {
931     jl[i] = mbs; il[0] = 0;
932   }
933   ierr = PetscMalloc(8*sizeof(MatScalar),&dk);CHKERRQ(ierr);
934   uik  = dk + 4;
935 
936   ai = a->i; aj = a->j; aa = a->a;
937 
938   /* for each row k */
939   for (k = 0; k<mbs; k++){
940 
941     /*initialize k-th row with elements nonzero in row k of A */
942     jmin = ai[k]; jmax = ai[k+1];
943     ap = aa + jmin*4;
944     for (j = jmin; j < jmax; j++){
945       vj = aj[j];         /* block col. index */
946       rtmp_ptr = rtmp + vj*4;
947       for (i=0; i<4; i++) *rtmp_ptr++ = *ap++;
948     }
949 
950     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
951     ierr = PetscMemcpy(dk,rtmp+k*4,4*sizeof(MatScalar));CHKERRQ(ierr);
952     i = jl[k]; /* first row to be added to k_th row  */
953 
954     while (i < k){
955       nexti = jl[i]; /* next row to be added to k_th row */
956 
957       /* compute multiplier */
958       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
959 
960       /* uik = -inv(Di)*U_bar(i,k): - ba[ili]*ba[i] */
961       diag = ba + i*4;
962       u    = ba + ili*4;
963       uik[0] = -(diag[0]*u[0] + diag[2]*u[1]);
964       uik[1] = -(diag[1]*u[0] + diag[3]*u[1]);
965       uik[2] = -(diag[0]*u[2] + diag[2]*u[3]);
966       uik[3] = -(diag[1]*u[2] + diag[3]*u[3]);
967 
968       /* update D(k) += -U(i,k)^T * U_bar(i,k): dk += uik*ba[ili] */
969       dk[0] += uik[0]*u[0] + uik[1]*u[1];
970       dk[1] += uik[2]*u[0] + uik[3]*u[1];
971       dk[2] += uik[0]*u[2] + uik[1]*u[3];
972       dk[3] += uik[2]*u[2] + uik[3]*u[3];
973 
974       /* update -U(i,k): ba[ili] = uik */
975       ierr = PetscMemcpy(ba+ili*4,uik,4*sizeof(MatScalar));CHKERRQ(ierr);
976 
977       /* add multiple of row i to k-th row ... */
978       jmin = ili + 1; jmax = bi[i+1];
979       if (jmin < jmax){
980         for (j=jmin; j<jmax; j++) {
981           /* rtmp += -U(i,k)^T * U_bar(i,j): rtmp[bj[j]] += uik*ba[j]; */
982           rtmp_ptr = rtmp + bj[j]*4;
983           u = ba + j*4;
984           rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1];
985           rtmp_ptr[1] += uik[2]*u[0] + uik[3]*u[1];
986           rtmp_ptr[2] += uik[0]*u[2] + uik[1]*u[3];
987           rtmp_ptr[3] += uik[2]*u[2] + uik[3]*u[3];
988         }
989 
990         /* ... add i to row list for next nonzero entry */
991         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
992         j     = bj[jmin];
993         jl[i] = jl[j]; jl[j] = i; /* update jl */
994       }
995       i = nexti;
996     }
997 
998     /* save nonzero entries in k-th row of U ... */
999 
1000     /* invert diagonal block */
1001     diag = ba+k*4;
1002     ierr = PetscMemcpy(diag,dk,4*sizeof(MatScalar));CHKERRQ(ierr);
1003     ierr = Kernel_A_gets_inverse_A_2(diag);CHKERRQ(ierr);
1004 
1005     jmin = bi[k]; jmax = bi[k+1];
1006     if (jmin < jmax) {
1007       for (j=jmin; j<jmax; j++){
1008          vj = bj[j];           /* block col. index of U */
1009          u   = ba + j*4;
1010          rtmp_ptr = rtmp + vj*4;
1011          for (k1=0; k1<4; k1++){
1012            *u++        = *rtmp_ptr;
1013            *rtmp_ptr++ = 0.0;
1014          }
1015       }
1016 
1017       /* ... add k to row list for first nonzero entry in k-th row */
1018       il[k] = jmin;
1019       i     = bj[jmin];
1020       jl[k] = jl[i]; jl[i] = k;
1021     }
1022   }
1023 
1024   ierr = PetscFree(rtmp);CHKERRQ(ierr);
1025   ierr = PetscFree(il);CHKERRQ(ierr);
1026   ierr = PetscFree(dk);CHKERRQ(ierr);
1027 
1028   C->factor    = FACTOR_CHOLESKY;
1029   C->assembled = PETSC_TRUE;
1030   C->preallocated = PETSC_TRUE;
1031   PetscLogFlops(1.3333*8*b->mbs); /* from inverting diagonal blocks */
1032   PetscFunctionReturn(0);
1033 }
1034 
1035 /*
1036     Numeric U^T*D*U factorization for SBAIJ format. Modified from SNF of YSMP.
1037     Version for blocks are 1 by 1.
1038 */
1039 #undef __FUNCT__
1040 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_1"
1041 int MatCholeskyFactorNumeric_SeqSBAIJ_1(Mat A,Mat *B)
1042 {
1043   Mat                C = *B;
1044   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
1045   IS                 ip = b->row;
1046   int                *rip,ierr,i,j,mbs = a->mbs,*bi = b->i,*bj = b->j;
1047   int                *ai,*aj,*r;
1048   int                k,jmin,jmax,*jl,*il,vj,nexti,ili;
1049   MatScalar          *rtmp;
1050   MatScalar          *ba = b->a,*aa,ak;
1051   MatScalar          dk,uikdi;
1052 
1053   PetscFunctionBegin;
1054   ierr  = ISGetIndices(ip,&rip);CHKERRQ(ierr);
1055   if (!a->permute){
1056     ai = a->i; aj = a->j; aa = a->a;
1057   } else {
1058     ai = a->inew; aj = a->jnew;
1059     ierr = PetscMalloc(ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr);
1060     ierr = PetscMemcpy(aa,a->a,ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr);
1061     ierr = PetscMalloc(ai[mbs]*sizeof(int),&r);CHKERRQ(ierr);
1062     ierr= PetscMemcpy(r,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr);
1063 
1064     jmin = ai[0]; jmax = ai[mbs];
1065     for (j=jmin; j<jmax; j++){
1066       while (r[j] != j){
1067         k = r[j]; r[j] = r[k]; r[k] = k;
1068         ak = aa[k]; aa[k] = aa[j]; aa[j] = ak;
1069       }
1070     }
1071     ierr = PetscFree(r);CHKERRQ(ierr);
1072   }
1073 
1074   /* initialization */
1075   /* il and jl record the first nonzero element in each row of the accessing
1076      window U(0:k, k:mbs-1).
1077      jl:    list of rows to be added to uneliminated rows
1078             i>= k: jl(i) is the first row to be added to row i
1079             i<  k: jl(i) is the row following row i in some list of rows
1080             jl(i) = mbs indicates the end of a list
1081      il(i): points to the first nonzero element in columns k,...,mbs-1 of
1082             row i of U */
1083   ierr = PetscMalloc(mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
1084   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
1085   jl   = il + mbs;
1086   for (i=0; i<mbs; i++) {
1087     rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0;
1088   }
1089 
1090   /* for each row k */
1091   for (k = 0; k<mbs; k++){
1092 
1093     /*initialize k-th row with elements nonzero in row perm(k) of A */
1094     jmin = ai[rip[k]]; jmax = ai[rip[k]+1];
1095 
1096     for (j = jmin; j < jmax; j++){
1097       vj = rip[aj[j]];
1098       rtmp[vj] = aa[j];
1099     }
1100 
1101     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
1102     dk = rtmp[k];
1103     i = jl[k]; /* first row to be added to k_th row  */
1104 
1105     while (i < k){
1106       nexti = jl[i]; /* next row to be added to k_th row */
1107 
1108       /* compute multiplier, update D(k) and U(i,k) */
1109       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
1110       uikdi = - ba[ili]*ba[i];
1111       dk += uikdi*ba[ili];
1112       ba[ili] = uikdi; /* -U(i,k) */
1113 
1114       /* add multiple of row i to k-th row ... */
1115       jmin = ili + 1; jmax = bi[i+1];
1116       if (jmin < jmax){
1117         for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j];
1118         /* ... add i to row list for next nonzero entry */
1119         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
1120         j     = bj[jmin];
1121         jl[i] = jl[j]; jl[j] = i; /* update jl */
1122       }
1123       i = nexti;
1124     }
1125 
1126     /* check for zero pivot and save diagoanl element */
1127     if (dk == 0.0){
1128       SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot");
1129       /*
1130     } else if (PetscRealPart(dk) < 0.0){
1131       SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Negative pivot: d[%d] = %g\n",k,dk);
1132       */
1133     }
1134 
1135     /* save nonzero entries in k-th row of U ... */
1136     ba[k] = 1.0/dk;
1137     jmin = bi[k]; jmax = bi[k+1];
1138     if (jmin < jmax) {
1139       for (j=jmin; j<jmax; j++){
1140          vj = bj[j]; ba[j] = rtmp[vj]; rtmp[vj] = 0.0;
1141       }
1142       /* ... add k to row list for first nonzero entry in k-th row */
1143       il[k] = jmin;
1144       i     = bj[jmin];
1145       jl[k] = jl[i]; jl[i] = k;
1146     }
1147   }
1148 
1149   ierr = PetscFree(rtmp);CHKERRQ(ierr);
1150   ierr = PetscFree(il);CHKERRQ(ierr);
1151   if (a->permute){
1152     ierr = PetscFree(aa);CHKERRQ(ierr);
1153   }
1154 
1155   ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr);
1156   C->factor    = FACTOR_CHOLESKY;
1157   C->assembled = PETSC_TRUE;
1158   C->preallocated = PETSC_TRUE;
1159   PetscLogFlops(b->mbs);
1160   PetscFunctionReturn(0);
1161 }
1162 
1163 /*
1164   Version for when blocks are 1 by 1 Using natural ordering
1165 */
1166 #undef __FUNCT__
1167 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering"
1168 int MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering(Mat A,Mat *B)
1169 {
1170   Mat                C = *B;
1171   Mat_SeqSBAIJ       *a=(Mat_SeqSBAIJ*)A->data,*b=(Mat_SeqSBAIJ *)C->data;
1172   int                ierr,i,j,mbs = a->mbs;
1173   int                *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j;
1174   int                k,jmin,*jl,*il,nexti,ili,*acol,*bcol,nz,ndamp = 0;
1175   MatScalar          *rtmp,*ba=b->a,*aa=a->a,dk,uikdi,*aval,*bval;
1176   PetscReal          damping=b->factor_damping, zeropivot=b->factor_zeropivot,shift_amount;
1177   PetscTruth         damp,chshift;
1178   int                nshift=0;
1179 
1180   PetscFunctionBegin;
1181   /* initialization */
1182   /* il and jl record the first nonzero element in each row of the accessing
1183      window U(0:k, k:mbs-1).
1184      jl:    list of rows to be added to uneliminated rows
1185             i>= k: jl(i) is the first row to be added to row i
1186             i<  k: jl(i) is the row following row i in some list of rows
1187             jl(i) = mbs indicates the end of a list
1188      il(i): points to the first nonzero element in U(i,k:mbs-1)
1189   */
1190   ierr = PetscMalloc(mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
1191   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
1192   jl   = il + mbs;
1193 
1194   shift_amount = 0;
1195   do {
1196     damp = PETSC_FALSE;
1197     chshift = PETSC_FALSE;
1198     for (i=0; i<mbs; i++) {
1199       rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0;
1200     }
1201 
1202     for (k = 0; k<mbs; k++){ /* row k */
1203     /*initialize k-th row with elements nonzero in row perm(k) of A */
1204       nz   = ai[k+1] - ai[k];
1205       acol = aj + ai[k];
1206       aval = aa + ai[k];
1207       bval = ba + bi[k];
1208       while (nz -- ){
1209         rtmp[*acol++] = *aval++;
1210         *bval++       = 0.0; /* for in-place factorization */
1211       }
1212       /* damp the diagonal of the matrix */
1213       if (ndamp||nshift) rtmp[k] += damping+shift_amount;
1214 
1215       /* modify k-th row by adding in those rows i with U(i,k) != 0 */
1216       dk = rtmp[k];
1217       i  = jl[k]; /* first row to be added to k_th row  */
1218 
1219       while (i < k){
1220         nexti = jl[i]; /* next row to be added to k_th row */
1221 
1222         /* compute multiplier, update D(k) and U(i,k) */
1223         ili   = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
1224         uikdi = - ba[ili]*ba[bi[i]];
1225         dk   += uikdi*ba[ili];
1226         ba[ili] = uikdi; /* -U(i,k) */
1227 
1228         /* add multiple of row i to k-th row ... */
1229         jmin = ili + 1;
1230         nz   = bi[i+1] - jmin;
1231         if (nz > 0){
1232           bcol = bj + jmin;
1233           bval = ba + jmin;
1234           while (nz --) rtmp[*bcol++] += uikdi*(*bval++);
1235           /* ... add i to row list for next nonzero entry */
1236           il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
1237           j     = bj[jmin];
1238           jl[i] = jl[j]; jl[j] = i; /* update jl */
1239         }
1240         i = nexti;
1241       }
1242 
1243       if (PetscRealPart(dk) < zeropivot && b->factor_shift){
1244 	/* calculate a shift that would make this row diagonally dominant */
1245 	PetscReal rs = PetscAbs(PetscRealPart(dk));
1246 	jmin      = bi[k]+1;
1247 	nz        = bi[k+1] - jmin;
1248 	if (nz){
1249 	  bcol = bj + jmin;
1250 	  bval = ba + jmin;
1251 	  while (nz--){
1252 	    rs += PetscAbsScalar(rtmp[*bcol++]);
1253 	  }
1254 	}
1255 	/* if this shift is less than the previous, just up the previous
1256 	   one by a bit */
1257 	shift_amount = PetscMax(rs,1.1*shift_amount);
1258 	chshift  = PETSC_TRUE;
1259 	/* Unlike in the ILU case there is no exit condition on nshift:
1260 	   we increase the shift until it converges. There is no guarantee that
1261 	   this algorithm converges faster or slower, or is better or worse
1262 	   than the ILU algorithm. */
1263 	nshift++;
1264 	break;
1265       }
1266       if (PetscRealPart(dk) < zeropivot){
1267         if (damping == (PetscReal) PETSC_DECIDE) damping = -PetscRealPart(dk)/(k+1);
1268         if (damping > 0.0) {
1269           if (ndamp) damping *= 2.0;
1270           damp = PETSC_TRUE;
1271           ndamp++;
1272           break;
1273         } else if (PetscAbsScalar(dk) < zeropivot){
1274           SETERRQ3(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %d value %g tolerance %g",k,PetscRealPart(dk),zeropivot);
1275         } else {
1276           PetscLogInfo((PetscObject)A,"Negative pivot %g in row %d of Cholesky factorization\n",PetscRealPart(dk),k);
1277         }
1278       }
1279 
1280       /* save nonzero entries in k-th row of U ... */
1281       /* printf("%d, dk: %g, 1/dk: %g\n",k,dk,1/dk); */
1282       ba[bi[k]] = 1.0/dk;
1283       jmin      = bi[k]+1;
1284       nz        = bi[k+1] - jmin;
1285       if (nz){
1286         bcol = bj + jmin;
1287         bval = ba + jmin;
1288         while (nz--){
1289           *bval++       = rtmp[*bcol];
1290           rtmp[*bcol++] = 0.0;
1291         }
1292         /* ... add k to row list for first nonzero entry in k-th row */
1293         il[k] = jmin;
1294         i     = bj[jmin];
1295         jl[k] = jl[i]; jl[i] = k;
1296       }
1297     } /* end of for (k = 0; k<mbs; k++) */
1298   } while (damp||chshift);
1299   ierr = PetscFree(rtmp);CHKERRQ(ierr);
1300   ierr = PetscFree(il);CHKERRQ(ierr);
1301 
1302   C->factor       = FACTOR_CHOLESKY;
1303   C->assembled    = PETSC_TRUE;
1304   C->preallocated = PETSC_TRUE;
1305   PetscLogFlops(b->mbs);
1306   if (ndamp) {
1307     PetscLogInfo(0,"MatCholeskyFactorNumerical_SeqSBAIJ_1_NaturalOrdering: number of damping tries %d damping value %g\n",ndamp,damping);
1308   }
1309  if (nshift) {
1310     PetscLogInfo(0,"MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering diagonal shifted %d shifts\n",nshift);
1311   }
1312 
1313   PetscFunctionReturn(0);
1314 }
1315 
1316 #undef __FUNCT__
1317 #define __FUNCT__ "MatCholeskyFactor_SeqSBAIJ"
1318 int MatCholeskyFactor_SeqSBAIJ(Mat A,IS perm,MatFactorInfo *info)
1319 {
1320   int ierr;
1321   Mat C;
1322 
1323   PetscFunctionBegin;
1324   ierr = MatCholeskyFactorSymbolic(A,perm,info,&C);CHKERRQ(ierr);
1325   ierr = MatCholeskyFactorNumeric(A,&C);CHKERRQ(ierr);
1326   ierr = MatHeaderCopy(A,C);CHKERRQ(ierr);
1327   PetscFunctionReturn(0);
1328 }
1329 
1330 
1331